A Hierarchical Multi-Resolution Self-Supervised Framework for High-Fidelity 3D Face Reconstruction Using Learnable...

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• 3D face reconstruction from single images is difficult due to ambiguous depth and entangled textures.
• Existing methods struggle with fine-scale details and robustness to variations.

Purpose of the Study:

• To propose a hierarchical multi-resolution self-supervised framework (HMR-Framework) for high-fidelity 3D face reconstruction.
• To progressively reconstruct coarse, medium, and fine-scale facial geometry within a unified pipeline.
• To enhance fine-scale texture fidelity and preserve edge-aware features.

Main Methods:

• Hierarchical multi-resolution self-supervised framework (HMR-Framework).
• Coarse geometric prior estimation via 3D morphable model regression.
• Medium-scale refinement using vertex deformation maps with a global-local Markov random field loss.
• Learnable Gabor-aware texture enhancement module for spatial-frequency decoupling.
• Wavelet-based detail perception loss for edge-aware texture preservation.

Main Results:

• The HMR-Framework achieves superior fine-detail reconstruction compared to state-of-the-art methods.
• The framework demonstrates robustness across various pose variations.
• Hierarchical design enhances semantic consistency across geometric scales.

Conclusions:

• The proposed HMR-Framework provides a functional solution for high-fidelity 3D face reconstruction from monocular images.
• The method effectively addresses challenges of depth ambiguity and multi-scale texture entanglement.
• The hierarchical approach improves accuracy and robustness in 3D face modeling.